Salient region detection using a discriminative color transform dictionary

Abstract

Salient region detection is a process of extracting a visually attractive region from a single image. In this paper, we present a novel method for automatically detecting salient region with discriminative color transform dictionary. The key assumption of the proposed framework is that the saliency map can be represented as a linear combination of color components of an image. To find a set of optimal coefficients for this linear combination, we incorporate least squares with constructing color transform dictionary, which includes multiple powers of color components in RGB and CIELab color space. In addition, we refine the saliency map by solving sparse representation problem with local dictionary, which is intended to enhance the compactness of the salient region in the saliency map. Experimental results show that the proposed method provides the saliency maps of two benchmark datasets with better quality and improved performance, compared to the previous state-of-the-art techniques.